DxOMark Sensor Performance: Nikon vs. Canon

[Orangutan]

I don't believe it's correct to say that 4 are combined into 1.  The de-mosaicing is an interpolation algorithm that uses the values from adjacent pixels to estimate the intensity and color of each pixel.  The demosaicing algorithm doesn't do any binning in the way you suggest. 

I didn't mean to suggest binning at all; demosaicing is (presumably) a complex algorithm.  I was making a simple point: if you take data from 4 photosites, each having 14EV of DR, you can demosaic in a way that gets you more than 14EV DR in the resultant pixel.  It's not guaranteed that you would, but it's possible. 

[Meh]

I don't believe it's correct to say that 4 are combined into 1.  The de-mosaicing is an interpolation algorithm that uses the values from adjacent pixels to estimate the intensity and color of each pixel.  The demosaicing algorithm doesn't do any binning in the way you suggest. 

I didn't mean to suggest binning at all; demosaicing is (presumably) a complex algorithm.  I was making a simple point: if you take data from 4 photosites, each having 14EV of DR, you can demosaic in a way that gets you more than 14EV DR in the resultant pixel.  It's not guaranteed that you would, but it's possible.

I was addressing only the first line of your statement where you said about demosaicing that 4 pixels are made into one RGB pixel which is not correct the way you wrote but fair enough if that's not what you meant.

Binning pixels still doesn't result in greater bit-depth though because that is limited by the ADC.  Each pixel can only be assigned 14-bit luminosity regardless of how many electrons are in the well.  For example some FF sensors have photosites with say 50,000 electron full-well capacity.  14-bit addressing only requires about 16,000 electrons so you have about 3 electrons per level.  Binning 4 pixels gives you more pixels but you still only have 14-bit addressing in the ADC.

But I'll leave it there since it's been a long time since I've been involved with that kind of thing and may already not be explaining it quite right.

[LetTheRightLensIn]

Hi,
   IMHO, DxOMark is not testing the sensor, but the camera itself because they use the RAW file produce by the DSLR (they didn't remove the sensor from the body and test, right?) and if the RAW file produce by the camera is not the RAW data, then it's not testing the sensor and the sensor compare between brands is somehow not valid. I believe all RAW file is not RAW data and Nikon RAW file are well know for it.

   Anyway, I was wondering how do Nikon D800 and D600 get a dynamic range of more than 14EV when there is only 14-bit of data? Compress the dynamic range or non-linear data?

   Have a nice day.

1. Why would they test the sensor removed from the camera? Can you get a picture out of a bare sensor? It's the RAW file that the user gets to use. Most of DSLR RAWs are pretty RAW under standard conditions and settings (Nikon does cook books starting with longer, but not long enough IMO, exposures, etc.).

2. They get more than 14EV on the normalized comparison, it's normalized to 8MP so with 36MP to play with there is a lot of room to get the extra bits after normalization there.

[sarangiman]

2. They get more than 14EV on the normalized comparison, it's normalized to 8MP so with 36MP to play with there is a lot of room to get the extra bits after normalization there.

Exactly. Say you have 14EV DR at the pixel level; i.e. you've perfectly matched your ADC to the DR attainable at the pixel level (e.g. saturation capacity = 16,384 e- & read noise = 1 e-). But say this is for a 100MP sensor.

When you go to make a 4" x 6" print of that 100MP file, you'll be doing a lot of downsizing for a 300dpi printer. That process of downsizing will reduce noise. Therefore you will be able to lift shadows more while still producing an acceptable print. Therefore shadows with a pretty low SNR may suddenly become 'acceptable' the viewer of the 4x6 print.

Yada yada... that's how you get more dynamic range via normalization.

Although I still have a hard time wrapping my head around it.

Because if your SNR is ever <1, you're not going to magically recover the signal by resampling... resampling would only bring you closer to the average of the noise as SNR approaches 0.

But for more reasonable definitions of DR (e.g. Bill Claff's 'PDR' that defines the lowest acceptable SNR as 20), normalization expanding DR makes perfect sense.

[dilbert]

Sensor performance isn't everything... but, if I were to choose Nikon or Canon today, I wouldn't be choosing Canon.

Someone I work with today was wondering what camera to buy and the first word out of my mouth was "D600". When they said that was too expensive, I said "D3200". There's just no way I can recommend Canon gear to anyone and feel good about it.

[dilbert]

...
Furthermore, their use case scores are normalized to an 8 MP file size, which explains how a camera with 14 bits per pixel can, according to DxOMark, actually deliver a dynamic range greater than 14 bits of EV. 
...

That depends on how the ADC works.

The traditional logic has been to squeeze lots of dark colours into a small space:
http://www.luminous-landscape.com/tutorials/optimizing_exposure.shtml
and that one bit of exposure (stop) is one bit position in the pixel. This provides a non-linear mapping of colour.

If they're using a different algorithm then it stands to reason that they're not limited by the number of bits in the colour palette.

[dilbert]

...
That said, DxO is a complete and total joke. This is a company that ranks consumer DSLRs above medium format digital backs in IQ.
...

There is no rule that says a MF digital back must be better at taking pictures than a DSLR.

If the design and technology used by the DSLR is superior to that of the digital back and delivers better images then it stands to reason that the sensor can be rated above a MF back.

[messus]

Canon has some catching up to do with respect to sensor performance as measured by http://www.DxOMark.com. Canon doesn't even come close to the top performing Nikons.  (High score is better.):

Pts Model
=======
96 Nikon D800E
95 Nikon D800
94 Nikon D600
81 Canon 5D III
79 Canon 5D II

(The Canon 1Dx is not yet rated.)
What are the chances that one of the reasons for the new sensor in the 6D is to catapult Canon's sensor performance into the mid 90's? I can't see Canon doing that considering the $3,500 EOS 5D III just came out and has a score of just 81. But Nikon's new $2,100 D600 kicks butt with a score of 94!

Sensor performance isn't everything... but, if I were to choose Nikon or Canon today, I wouldn't be choosing Canon.

I have said this many times. You CAN NOT compare dynamic range on ISO100, and assume everybody works in ISO100, and that this is the correct ISO to measure by, like DxO does. I RARELY go below ISO 1600 for my kind of work.

BTW! Anyone find it strange DxO has the finished analysis/review of the D600 days after it is released? And they still haven't finished they analysis/review of the 1D-X !?

Well I don't find it strange at all, DxO is tight with Nikon, everybody should know that by now. I will not be surprised if they rate the 1D-X below the Nikon D3100, and neither should you! Do not expose your low intelligence by actually believing DxO is the standard for measuring camera low light and dynamic range performance.

The 5D3 has a MUCH better dynamic range, not to mention the low light ISO performance, than the D800 above ISO600.

Who cares what the Nikonioans down at DxO says, not me!

Have a nice day everyone!

[dilbert]

...
I have said this many times. You CAN NOT compare dynamic range on ISO100, and assume everybody works in ISO100, and that this is the correct ISO to measure by, like DxO does. I RARELY go below ISO 1600 for my kind of work.

More than 90% of the photos I take are at ISO100, if only because that's what gives me the best IQ.

BTW! Anyone find it strange DxO has the finished analysis/review of the D600 days after it is released? And they still haven't finished they analysis/review of the 1D-X !?
...

Nope.

Nikon had the D600 on sale via the web/stores 5 days after it was announced, so it is likely that Nikon were able to ship a production matched D600 to DxO ahead of time.

How long did it take Canon to even get 1DX cameras out to folks that were going to need them for the Olympics? And how long did it take for the 1DX to show up in stores? Don't blame DxO for Canon's inability to delivery/perform.

[mystic_theory]

It's amazing how much distance in DxO scores there is between Nikon sensors and the corresponding Canon ones: I agree that the overall score gives just a rough estimate of the overall quality of the sensor, but what worries me are the differences in high ISO noise and DR, which are instead very precise and reliable measurements.   

[sarangiman]

...
That said, DxO is a complete and total joke. This is a company that ranks consumer DSLRs above medium format digital backs in IQ.
...

There is no rule that says a MF digital back must be better at taking pictures than a DSLR.

If the design and technology used by the DSLR is superior to that of the digital back and delivers better images then it stands to reason that the sensor can be rated above a MF back.

+1

Thank you dilbert for stating what with common sense & logic should be apparent to everyone.

[sarangiman]

I have said this many times. You CAN NOT compare dynamic range on ISO100, and assume everybody works in ISO100, and that this is the correct ISO to measure by, like DxO does. I RARELY go below ISO 1600 for my kind of work.

You do realize that at base ISO, the D800 has more than 2.5EV more DR than the 5DIII, and at worse, i.e. at ISO 25,600, it has 0.26EV worse DR than the 5DIII. So, compare at whatever ISO you'd like. The D800 is rarely going to do worse than the 5DIII, but can do much better.

Furthermore, the future is likely 'ISO-less'. In other words, ISO becomes metadata written into the file, much like white balance. Amplification is done during RAW processing. That way, you actually retain the DR of base ISO even at high ISO (w/ low enough read noises).

We just need manufacturers (both hardware & software) to implement this. Personally, I would've already implemented this w/ the D800, given the low read noise of ~3 e-.

But maybe the industry is waiting for a read noise of 1 e-, or a 16-bit ADC.

[Fishnose]

...
That said, DxO is a complete and total joke. This is a company that ranks consumer DSLRs above medium format digital backs in IQ.
...

There is no rule that says a MF digital back must be better at taking pictures than a DSLR.

If the design and technology used by the DSLR is superior to that of the digital back and delivers better images then it stands to reason that the sensor can be rated above a MF back.

+2

[jthomson]

BTW! Anyone find it strange DxO has the finished analysis/review of the D600 days after it is released? And they still haven't finished they analysis/review of the 1D-X !?
...

Nope.

Nikon had the D600 on sale via the web/stores 5 days after it was announced, so it is likely that Nikon were able to ship a production matched D600 to DxO ahead of time.

How long did it take Canon to even get 1DX cameras out to folks that were going to need them for the Olympics? And how long did it take for the 1DX to show up in stores? Don't blame DxO for Canon's inability to delivery/perform.
[/quote]

DXO appears to have a Nikon bias.  They haven't released scores for the 1DX, the T4i or the Olympus OMD EM-5, but they get the D600 out within days of its release. 

[dilbert]

BTW! Anyone find it strange DxO has the finished analysis/review of the D600 days after it is released? And they still haven't finished they analysis/review of the 1D-X !?
...

Nope.

Nikon had the D600 on sale via the web/stores 5 days after it was announced, so it is likely that Nikon were able to ship a production matched D600 to DxO ahead of time.

How long did it take Canon to even get 1DX cameras out to folks that were going to need them for the Olympics? And how long did it take for the 1DX to show up in stores? Don't blame DxO for Canon's inability to delivery/perform.

DXO appears to have a Nikon bias.  They haven't released scores for the 1DX, the T4i or the Olympus OMD EM-5, but they get the D600 out within days of its release.

Maybe Nikon sent DxO a D600 free of charge and ahead of time because they wanted the results up on the DxO web page ASAP?

Why would Nikon do that?

Because Nikon are smart enough to realise that they will score well on DxO and that it'll be good marketing and in general a boost for their brand and sales.

If you were DxO and Nikon did such, would you say no?

The rub here is that if you're Canon and you already know that your camera's sensor isn't going to score better than anything Nikon has published then you're not going to be in any sort of rush to have it tested by DxO. I don't know if DxO purchases the cameras for themselves or waits for vendors to send them freebies...